Publication number | US6961719 B1 |
Publication type | Grant |
Application number | US 10/043,044 |
Publication date | Nov 1, 2005 |
Filing date | Jan 7, 2002 |
Priority date | Jan 7, 2002 |
Fee status | Lapsed |
Also published as | US7293001 |
Publication number | 043044, 10043044, US 6961719 B1, US 6961719B1, US-B1-6961719, US6961719 B1, US6961719B1 |
Inventors | Man Mohan Rai |
Original Assignee | The United States Of America As Represented By The Administrator Of The National Aeronautics And Space Administration |
Export Citation | BiBTeX, EndNote, RefMan |
Patent Citations (17), Non-Patent Citations (5), Referenced by (15), Classifications (13), Legal Events (7) | |
External Links: USPTO, USPTO Assignment, Espacenet | |
The invention disclosed herein was made by an employee of the United States Government and may be manufactured and used by or for the Government for governmental purposes without payment of any royalties for such manufacture and use.
This invention relates to design optimization, using a hybrid neural network and support vector machine approach to construct a response surface that models a selected objective function.
Considerable advances have been made in the past two decades in developing advanced techniques for numerical simulation of fluid flows in aerodynamic configurations. These techniques are now mature enough to be used routinely, in conjunction with experimental results, in aerodynamic design. However, aerodynamic design optimization procedures that make efficient use of these advanced techniques are still being developed.
The design of aircraft components, such as a wing, a fuselage or an engine, involves obtaining an optimal component shape that can deliver the desired level of component performance, subject to one or more constraints (such as maximum weight or cost) that the component(s) must satisfy. Aerodynamic design can be formulated as an optimization problem that requires minimization of an objective function, subject to constraints. Many formal optimization methods have been developed and applied to aerodynamic design. These include inverse design methods, adjoint methods, sensitivity derivative-based methods and traditional response surface methodology (RSM).
Inverse design methods in aerodynamics are used to provide a component that responds in a pre-selected manner, for example, an aircraft wing that has a prescribed pressure distribution. The known inverse methods do not account for certain fluid parameters, such as viscosity, and are used in preliminary design only.
Adjoint methods provide a designer with the gradient of the objective function. One advantage of this method is that the gradient information is obtained very quickly. However, where several technical disciplines are applied simultaneously, it is often difficult to perform design optimization using this method; each discipline requires a different formulation. It is also difficult and expensive to quickly evaluate the effects of engineering tradeoffs, where the applicable constraints may be changed several times. It is also not possible to use existing experimental data or partial or unstructured data in the design process.
A sensitivity derivative-based method typically requires that a multiplicity of solutions, with one parameter varied at a time, be obtained to compute a gradient of the objective function. The number of computations required grows linearly with the number of design parameters considered for optimization, and this method quickly becomes computationally expensive. This method is also sensitive to noise present in the design data sets. As with an adjoint method, it is not possible to use existing experimental data or partial or unstructured data in the design process.
RSM provides a framework for obtaining an optimal design, using statistical procedures, such as regression analysis and design of experiments. Traditional RSM uses low-degree regression polynomials in the relevant design variables to model the variation of an objective function. The polynomial model is then analyzed to obtain an optimal design. Several polynomial models may have to be constructed to provide an adequate view of the design space. Addition of higher degree polynomials will increase the computational cost and will build in higher sensitivity to noise in the data used.
Artificial neural networks (“neural nets” herein) have been widely used in fields such as aerodynamic engineering, for modeling and analysis of flow control, estimation of aerodynamic coefficients, grid generation and data interpolation. Neural nets have been used in RSM-based design optimization, to replace or complement a polynomial-based regression analysis. Current applications of neural nets are limited to simple designs involving only a few design parameters. The number of data sets required for adequate modeling may increase geometrically or exponentially with the number of design parameters examined. A neural net analysis requires that the design space be populated with sufficiently dense simulation and/or experimental data. Use of sparse data may result in an inaccurate representation of the objective function in design space. On the other hand, inefficient use of design data in populating the design space can result in excessive simulation costs. Capacity control is critical to obtain good generalization capability. In some preceding work, this problem was alleviated by using a neural net to represent the functional behavior with respect to only those variables that result in complex, as opposed to simple, variations of the objective function; the functional behavior of the remaining variables was modeled using low degree polynomials. This requires a priori knowledge to partition the design variables into two sets.
Use of neural network (NN) analysis of a physical object, in order to optimize response of the object in a specified physical environment, is well known. An example is optimization of a turbine blade shape, in two or three dimensions, in order to reproduce an idealized pressure distribution along the blade surface, as disclosed by Rai and Madavan in “Aerodynamic Design Using Neural Networks”, AIAA Jour., vol. 38 (2000) pp. 173–182. NN analysis is suitable for multidimensional interpolation of data that lack structure and provides a natural structure in which a succession of numerical solutions of increasing complexity, or increasing fidelity to a real world environment, can be represented and optimized. NN analysis is especially useful when multiple design objectives need to be met.
A feed-forward neural net is a nonlinear estimation technique. One difficulty associated with use of a feed-forward neural net arises from the need for nonlinear optimization to determine connection weights between input, intermediate and output variables. The training process can be very expensive when large amounts of data need to be modeled.
In response to this, a support vector machine (SVM) approach, originally applied in statistical learning theory, has been developed and applied. Support vector machine analysis allows use of a feature space with a large dimension, through use of a mapping from input space into feature space and use of a dual formulation of the governing equations and constraints. One advantage of an SVM approach is that the objective function (which is to be minimized to obtain the coefficients that define the SVM model) is convex so that any local minimum is also a global minimum; this is not true for many neural net models. However, an underlying feature space (polynomial, Gaussian, etc.) must be specified in a conventional SVM approach, and data resampling is required to implement model hybridization. Hybridization is more naturally, and less expensively, applied in a neural net analysis.
What is needed is a machine learning algorithm that combines the desirable features of NN analysis and of SVM analysis and does not require intimate a priori familiarity with operational details of the object to be optimized. Preferably, the method should automatically provide a characterization of many or all of the aspects in feature space needed for the analysis.
The invention meets these needs by providing a hybrid of NN analysis and SVM analysis, referred to as NN/SVM analysis herein. In one embodiment, NN/SVM analysis begins with a group of associated, independent input space coordinates (parameter values), maps these coordinates into a feature space of appropriately higher dimension that includes a computed set of combinations (e.g., powers) of the input space coordinates with the assistance of the input and hidden layers of an NN, constructs an inner product formalism for the coordinates in feature space, obtains a solution to a minimization problem to compute Lagrange multiplier values that define the SVM, and returns to input space to complete a solution of the problem.
FIGS. 8A/8B/8C are a flow chart for an RSM procedure used in practicing the invention.
FIGS. 9A1–9C2 graphically illustrate evolution of an airfoil and corresponding pressure distribution obtained from an iterative NN/SVM analysis.
FIGS. 10 and 11A/11B illustrate data classification in two dimensions.
Consider a feed-forward neural network 21 having an input layer with nodes 23-m (m=1, . . . , 5), a hidden layer with nodes 25-n (n=1, 2, 3), and an output node 26, as illustrated schematically in
of a weighted sum of the parameter values p_{m}, where C_{nm }is a connection weight, which can be positive, negative or zero, linking an input node 23-m with a hidden layer node 25-n. The output of the network 21 is assumed for simplicity, initially, to be a single-valued scalar,
It is known that NN approximations of the format set forth in Eqs. (1) and (2) are dense in the space of continuous functions when the activation functions Φ_{n }are continuous sigmoidal functions (monotonically increasing functions, with a selected lower limit, such as 0, and a selected upper limit, such as 1). Three commonly used sigmoidal functions are
Φ(z)=1/{1+exp(−z)}, (3A)
Φ(z)=(1+tanh(z)}/2, (3B)
Φ(z)={π+2·tan^{−1}(z)}/2π, (3C)
Other sigmoidal functions can also be used here. In the context of design optimization, a trained NN represents a response surface, and the NN output is the objective function. In multiple objective optimization, different NNs can be used for different objective functions. A rapid training algorithm that determines the connection weights C_{nm }and coefficients D_{n }is also needed here.
The approach set forth in the preceding does reasonably well in an interpolative mode, that is, in regions where data points (parameter value vectors) are reasonably plentiful. However, this approach rarely does well in an extrapolative mode. In this latter situation, a precipitous drop in estimation accuracy may occur as one moves beyond the convex hull defined by the data point locations. In part, this is because the sigmoidal functions are not the most appropriate basis functions for most data modeling situations. Where the underlying function(s) is a polynomial in the parameter values, a more appropriate set of basis functions is a set of Legendre functions (if the parameter value domain is finite), or a set of Laguerre or Hermite functions (if the parameter value domain is infinite). Where the underlying function(s) is periodic in a parameter value, a Fourier series may be more appropriate to represent the variation of the function with that parameter.
Two well known approaches are available for reducing the disparity between an underlying function and an activation function. A first approach, relies on neural nets and uses appropriate functions of the primary variables as additional input signals for the input nodes. These functions simplify relationships between neural net input and output variables but require a priori knowledge of these relationships, including specification of all the important nonlinear terms in the variables. For example, a function of the (independent) parameter values x and y, such as
h(x,y)=a·x ^{2} +b·x·y+c·y ^{2} +d·x+e·y+f, (5)
where a, b, c, d, e and f are constant coefficients, would be better approximated if the terms x, y, x^{2}, x·y and y^{2 }are all supplied to the input nodes of the network 21. However, in a more general setting with many parameters, this leads to a very large number of input nodes and as-yet-undetermined connection weights C_{nm}.
A second approach, referred to as a support vector machine (SVM), provides a nonlinear transformation from the input space variables p_{m }into a feature space that contains the original variables p_{m }and the important nonlinear combinations of such terms (e.g., (p_{1})^{2}, (p^{1})(p_{2})^{3}(p_{M})^{2 }and exp(p_{2})) as coordinates. For the example function h(p_{1},p_{2}) set forth in Eq. (5), the five appropriate feature space coordinates would be p_{1}, p_{2}, (p_{1})^{2}, p_{1}·p_{2 }and (p_{2})^{2}. Very high dimensional feature spaces can be handled efficiently using kernel functions for certain choices of feature space coordinates. The total mapping between the input space of individual variables (first power of each parameter p_{m}) and the output space is a hyperplane in feature space. For a model that requires only linear terms and polynomial terms of total degree 2 (as in Eq. (5)), in the input space variables, the model can be constructed efficiently using kernel functions that can be used to define inner products between vectors in feature space. However, use of an SVM requires a priori knowledge of the functional relationships between input and output variables.
The mapping between the input space parameters and the output function is defined using a kernel function and certain Lagrange multipliers. The Lagrange multipliers are obtained by maximizing a function that is quadratic and convex in the multipliers, the advantage being that every local minimum is also a global minimum. By contrast, a neural net often exhibits numerous local minima of the training error(s) that may not be global minima. However, several of these local minima may provide acceptable training errors. The resulting multiplicity of acceptable weight vectors can be used to provide superior network generalization, using a process known as network hybridization. A hybrid network can be constructed from the individual trained networks, without requiring data re-sampling or similar expensive techniques.
An attractive feature of a neural net, vis-a-vis an SVM, is that the coordinates used in a feature space do not have to be specified (e.g., via kernel functions). However, use of an SVM, in contrast to use of a neural net, allows one to introduce features spaces with a large number of dimensions, without a corresponding increase in the number of coefficients.
A primary contribution of the present invention is to provide a mechanism, within the NN component, for determining at least the coordinate (parameter) combinations needed to adequately define the feature space for an SVM, without requiring detailed knowledge of the relationships between input parameters and the output function.
In step 43, feature space inner products that are required for the SVM are computed. In step 43A, user-specified feature space coordinates and corresponding inner products and kernel functions are provided. Note that the feature space is a vector space with a corresponding inner product.
In step 44, a Lagrange functional is defined and minimized, subject to constraints, to obtain Lagrange multiplier values for the SVM. See the Appendix for a discussion of a Lagrange functional and associated constraints. In step 45, the NN connection weights and the Lagrange multiplier coefficients are incorporated and used to compute a training error associated with this choice of values within the NN/SVM.
In step 46, the system determines if the training error is no greater than a specified threshold level. If the answer to the query in step 46 is “no”, the system changes at least one connection weight, in step 47, preferably in a direction that is likely to reduce the training error, and repeats steps 42–46. If the answer to the query in step 46 is “yes”, the system interprets the present set of connection weights and Lagrange multiplier values as an optimal solution of the problem, in step 48.
Note that steps 42–48 can be embedded in an optimization loop, wherein the connection weights are changed according to the rules of the particular optimization method used.
The hybrid NN/SVM system relies on the following broadly stated actions: (1) provide initial random (or otherwise specified) connection weights for the NN; (2) use the activation function(s) and the connection weights associated with each hidden layer unit to construct inner products for the SVM; (3) use the inner products to compute the Lagrange multiplier values; (4) compute a training error associated with the present values of the connection weights and Lagrange multiplier values; (5) if the training error is too large, change at least one connection weight and repeat steps (2)–(4); (6) if the training error is not too large, accept the resulting values of the connection weights and the Lagrange multiplier values as optimal.
This method has several advantages over a conventional SVM approach. First, coordinates that must be specified a priori in the feature space for a conventional SVM are determined by the NN component in an NN/SVM system. The feature space coordinates are generated by the NN component to correspond to the data at hand. In other words, the feature space provided by the NN component evolves to match or correspond to the data. A feature space that evolves in this manner is referred to as “data-adaptive.” The feature space coordinates generated by the NN component can be easily augmented with additional user-specified feature space coordinates (parameter combinations) and kernel functions.
Second, use of activation functions that are nonlinear functions of the connection weights in the NN component reintroduces the possibility of multiple local minima and provides a possibility of hybridization without requiring data resampling.
The feature spaces generated by the NN hidden layer can be easily augmented with high-dimensional feature spaces without requiring a corresponding increase in the number of connection weights. For example, a polynomial kernel containing all monomials and binomials (degrees one and two) in the parameter space coordinates can be added to an inner product generated by the SVM component, without requiring any additional connection weights or Lagrange multiplier coefficients.
The NN/SVM system employs nonlinear optimization methods to obtain acceptable connection weights, but the weight vectors thus found are not necessarily unique. Many different weight vectors may provide acceptably low training errors for a given set of training data. This multiplicity of acceptable weight vectors can be used to advantage. If validation data are available, one can select the connection weight vector and resulting NN/SVM system with the smallest validation error. In aerodynamics, this requires additional simulations that can be computationally expensive.
If validation data are not available, multiple trained NNs or NN/SVM systems can be utilized by creating a hybrid NN/SVM. A weighted average of N output signals from trained NN/SVMs in a hybrid NN/SVM is formed as a new solution. Where the weights are equal, if errors for the N individual output solutions are uncorrelated and individually have zero mean, the least squares error of this new solution is approximately a factor of N less than the average of the least squares errors for the N individual solutions. When the errors for the N individual output solutions are partly correlated, the hybrid solution continues to produce a least squares error that is smaller than the average of the least squares errors for the N individual solutions, but the difference is not as large. The N trained NN/SVMs used to form a hybrid system need not have the same architecture or be trained using the same training set.
The generalization in each of
In step 82, initial values of the parameters, p=p0, are provided from an initial approximation to the desired airfoil shape.
In step 83, optimal data values P(r_{k};opt) (e.g., airfoil pressure values or airfoil heat transfer values) are provided at selected locations r_{k}=(x_{k},y_{k},z_{k}) (k=1, . . . , K) on the airfoil perimeter.
In step 84, an equilateral M-simplex, denoted MS(p0), is constructed, with a centroid or other selected central location at p=p0, in M-dimensional parameter space, with vertices lying on a unit radius sphere. Each of the M+1 vertices of the M-simplex MS(p0) is connected to the centroid, p=p0, by a vector Δp(m) (m=1, . . . , M+1) in parameter space. More than the M+1 vertices can be selected and used within the M-simplex. For example, midpoints of each of the M(M+1)/2 simplex edges can be added to the M+1 vertices. These additional locations will provide a more accurate NN/SVM model.
In step 85, a computational fluid dynamics (CFD) or other calculation is performed for an extended parameter value set, consisting of the parameter value vectors p=p0 and each of the M+1 M-simplex vertices, P=P_{vert}=P0+Δp(m), to obtain a calculated pressure distribution P(r_{k};p_{vert}) at each of the selected perimeter locations, r=r_{k }for each of these parameter value sets. One hybrid NN/SVM is assigned to perform the analysis for all vertices in the M-simplex MS(p0) at each location r_{k}. That is, a total of K NN/SVM systems are used to model the overall pressure dependence on the parameters p_{m}. The calculated pressure distribution P(r_{k};p_{vert}) and/or the airfoil can be replaced by any other suitable physical model, in aerodynamics or in any other technical field or discipline. Used together, the trained NN/SVM systems will provide the pressure distribution P(r_{k};p) for general parameter value vectors p.
In step 86, a first objective function, such as
is introduced, where {w_{k}} is a selected set of non-negative weight coefficients.
In step 87, the minimum value of the first objective function OBJ(p;p0;1) and a corresponding parameter vector p=p(min) are determined for parameter vectors p within a selected sphere having a selected diameter or dilatation factor d, defined by |p−p0|≦d, with 1<d≦10. The process is performed using a nonlinear optimization method. Other measures of extrapolation can also be used here.
In step 88, the system calculates a second objective function, which may be the first objective function or (preferably) may be defined as
where P(r_{k};p;CFD) is a pressure value computed using a CFD simulation, for p=p(min) and p=p0. The system then determines if OBJ(p(min);p0;2)<OBJ(p0;p0;2) for the intermediate minimum value parameter vector, p=p(min). One can use the first objective function OBJ(p;p0;1), defined in Eq. (6A), rather than the objective function OBJ(p;p0;2) defined in Eq. (6B), for this comparison, but the resulting inaccuracies may be large.
If the answer to the query in step 88 is “no” for the choice of dilatation factor d, the dilatation factor d is reduced to a smaller value d′ (1<d′<d), in step 89, and steps 88 and 89 are repeated until the approximation pressure values {P(r_{k},p)}_{k }for the extrapolated parameter value set provide an improved approximation for the optimal values for the same airfoil perimeter locations, r=r_{k}.
If the answer to the query in step 88 is “yes”, the system moves to step 90, uses the (modified) objective function and uses the intermediate minimum-cost parameter value set, p=p(min), which may lie inside or outside the M-simplex MS(p0) in parameter space. Minimization of the objective function OBJ(p;p0) may include one or more constraints, which may be enforced using the well known method of penalty functions. The (modified) objective function definition in Eq. (6A) (or in Eq. (6B)) can be replaced by any other positive definite definition of an objective function, for example, by
where q is a selected positive number.
If the original parameter value set p has an insufficient number of parameters, this will become evident in the preceding calculations, and the (modified) objective function OBJ(p(min);p0) or OBJ(p(min);p0)* will not tend toward acceptably small numbers. In this situation, at least one additional parameter would be added to the parameter value set p and the procedure would be repeated. In effect, an NN/SVM procedure used in an RSM analysis will require addition of (one or more) parameters until the convergence toward a minimum value that is acceptable for an optimized design.
In step 91, the system determines if the (modified) objective function OBJ(p(min);p0)* is no greater than a selected threshold number (e.g., 1 or 10^{−4}, in appropriate units). If the answer to the query in step 91 is “no”, a new M-simplex MS(p′0) is formulated, in step 92, with p′0=p(min) as the new center, and steps 85–90 are repeated at least once. Each time, a new parameter value set, p=p(min), is determined that approximately minimizes the objective function OBJ(p;p′0).
If the answer to the query in step 91 is “yes”, the system interprets the resulting parameter set, p=p(min), and the design described by this parameter set as optimal, in step 93. The method set forth in steps 81–93 is referred to herein as a response surface method.
FIGS. 9A1–9C2 illustrate a sequence of partly-optimized designs for an airfoil, obtained using the invention, and compare each such design shape and corresponding airfoil pressure distribution to an target airfoil design shape and corresponding target airfoil pressure distribution. The objective function is defined as mean square error between resulting and target pressure distribution at a sequence of selected locations on the airfoil perimeter. One begins in FIG. 9A1 with a curvilinear shape of approximately uniform thickness, which provides a pressure distribution p along the airfoil perimeter as illustrated graphically in FIG. 9A2. FIGS. 9B1 and 9C1 illustrate the results of second and fourth iterative applications of an NN/SVM analysis according to the invention, and FIGS. 9B2 and 9C2 graphically illustrate the pressure distributions corresponding to FIGS. 9B1 and 9C1, respectively. Each iteration brings the resulting airfoil shape and pressure distribution closer to the target shape and target pressure distribution. After a fourth iteration of the NN/SVM analysis, the airfoil shape, shown in FIG. 9C1, produces a pressure distribution, shown in FIG. 9C2, that nearly precisely matches the target airfoil pressure distribution. Computations for this iterative sequence required about 8 minutes on a 16-processor SGI Origin computer.
In a second embodiment, NN/SVM analysis is applied to data classification in a multi-dimensional vector space. In data classification, a discrimination mechanism must be determined that divides the data points into (at least) a first set of data points that satisfy a selected criterion, and a second set of data points that either do not satisfy the (first) criterion or that satisfy an inconsistent second criterion.
f _{1}(x,y)=a·x+b·y−c=0, (7)
where a, b and c are selected real values, with at least one of a and b being non-zero: All data points in the first data set and in the second data set lie on opposite sides of the line (hyperplane) f_{1}(x,y)=0. Here, the data point separation is straightforward.
f _{2}(x,y)=(a·x+b·y−c)^{2}±(d·x+e·y−g)^{2}=1, (8)
where a·d+b·e=0 and a, b, c, d, e and g are selected real values, not all zero. The choice of the plus (+) sign in Eq. (8) produces an ellipse, and the choice of a minus (−) sign in Eq. (8) produces a hyperbola. In this instance, one set of appropriate coordinates for hyperplane separation in feature space is
u_{1}=x, (9A)
u_{2}=y, (9B)
u _{3}=(a·x+b·y−c)^{2}, (9C)
u _{4}=(d·x+e·y−g)^{2}, (9D)
in which the separating hyperplane in feature space becomes
u _{3} ±u _{4}−1=0. (10)
The power of an SVM resides, in part, in its use of a qth order polynomial kernel (as an example) for vectors α and β, such as
K(α,β)=(α·β+1)^{q}, (11)
where q is a selected positive integer (e.g., q=2), rather than requiring an a priori definition of the polynomial terms to be used, as in Eqs. (9A)–(9D).
An advantage of the present invention, using NN/SVM analysis, over a conventional SVM analysis is that the kernel, such as the one given in Eq. (11), and the associated feature space need not be specified a priori; the appropriate feature space is automatically generated by the NN component of the NN/SVM system during the training process.
ρ={(x,y)|0≦x≦2.5,0≦y≦2.5}. (12)
Four hundred data points were randomly generated in this region and were first classified according to the exact boundaries. The boundaries were then removed, and only the locations of the data points were provided to the NN/SVM system. The resulting decision boundary generated by the NN/SVM system is shown as a solid line in
The NN/SVM system provides a perfect classification of the original data, with zero mis-assignments, without requiring any specification of kernel functions or feature spaces. Where the solid boundary line and the dotted boundary lines differ, no data points were located in the intervening regions between these boundaries. Provision of additional data points in one or more of these intervening regions would provide a resulting (solid) NN/SVM boundary line that is closer to the exact (dotted) boundary line.
If r is a ratio of the sum of the absolute value of the intervening regions corresponding to the boundary lines mismatch, and the area of the square (6.25 units^{2 }in
Examples of an NN analysis and of an SVM analysis are presented here. The invention is not limited to a particular NN analysis or to a particular SVM analysis.
Consider an object, represented by a group of coordinates x=(x^{1}, x^{2}, . . . , x^{N}), for which some physical feature or response of the object is to be optimized. The object may be a aircraft wing or turbine blade for which an ideal pressure distribution at specified locations on the object is to be achieved as closely as possible. The object may be a chemically reacting system with desired percentages of final compounds, for which total thermal energy output is minimized. The object may be represented at spaced apart locations or at spaced apart times by a group of independent coordinates, and an objective or cost function is presented, representing the response to be optimized. One or more constraints, either physical or numerical, are also set down, if desired.
In an NN analysis, one relevant problem is minimizing empirical risk over a sum of linear indicator or characteristic functions
where θ is an indicator or characteristic function, x is a coordinate vector and w is a vector of selected weight coefficients. Consider a training set of (N+1)-tuples (x_{1},y_{1}), (x_{2},y_{2}), . . . , (x_{K},y_{K}), where each x_{j}=(x_{j} ^{1}, x_{j} ^{2}, . . . ,x_{j} ^{N}) is an N-tuple representing a vector and y_{j }is a scalar having only the values −1 or +1.
The indicator function θ(z) has only two values, 0 and 1, and is not generally differentiable with respect to a variable in its argument. The indicator function θ(z) in Eq. (A-1) is often replaced by a general sigmoid function S(z) that is differentiable with respect to z everywhere on the finite real line, is monotonically increasing with z, and satisfies
Lim _{z→−∞} S(z)=0, (A-2a)
Lim _{z→+∞} S(z)=1. (A-2b)
Examples of suitable sigmoid functions include the following:
S(z)=1/{1+exp(−αz)},
S(z)={1+tanh(β·z+χ)]/2
S(z)={π+2·tan^{−1}(δ·z+ε}/2π,
where α, β and δ are selected positive values. The indicator sum f(x,w) in Eq. (A-1) is replaced by a modified sigmoid sum
where S is a selected linear or nonlinear function.
In order to minimize the empirical risk, one must determine the parameter values w_{i }that minimize an empirical risk functional
which is differentiable in the vector components w. One may, for example, use a gradient search approach to minimize R_{emp}(w). The search may converge to a local minimum, which may or may not be a global minimum for the empirical risk.
Assume, first, that the training data {(x_{j},y_{j})} can be separated by an optimal separating hyperplane, defined by
(w·x _{j})−g=0, (A-5)
where g partly defines the hyperplane. A separating hyperplane satisfies
(w·x _{j})−g≧1 (y _{j}≧1), (A-6a)
(w·x _{j})−g≦−1 (y _{j}≧−1), (A-6b)
An optimal separating hyperplane maximizes the functional
Φ(w)=(w·w)/2, (A-7)
with respect to the vector values w and the value g, subject to the constraints in Eqs. (A-6a)–(A-6b). Unless indicated otherwise, all sums in the following are understood to be over the index j (j=1, . . . , K).
A solution to this optimization problem is given by a saddle point of a Lagrange functional
At a saddle point, the solutions (w,g,α) satisfy the relations
∂L/∂g=0, (A-9)
∂L/∂w=0, (A-10)
with the associated constraint
αj≧0, (A-11)
Equation (A-9) yields the constraint
Equation (A-10) provides an expression for the parameter vector w of an optimal hyperplane as a linear combination of vectors in the training set
w=Σy _{j}·α_{j} ·x _{j}, (A-13)
An optimal solution (w,g,α) must satisfy a Kuhn-Tucker condition
αj{((x _{j} ·w)−g)·(y _{j}−1)=0 (=1, . . . , K). (A-14)
Only some of the training vectors, referred to herein as “support vectors,” have non-zero coefficients in the expansion of the optimal solution vector w. More precisely, the expansion in Eq. (A-13) can be rewritten as
w=Σy _{j}·α_{j} ·x _{j}. (A-15)
The preceding development assumes that the training set data {(x_{j},y_{j})} are separable by a hyperplane. If these data are not separable by a hyperplane, one introduces non-negative slack variables χ_{j}(j=1, . . . , K) and a modified functional
Φ(w)=(w·w)+C·Σχ _{j}, (A-18)
subject to the constraints
y _{j·}((w·x _{j})−g)≧1−x _{j} (A-19)
where the (positive) coefficient C corresponds to an inter-penetration of two or more groups of training set (N+1)-tuples into each other (thus, precluding separation by a hyperplane). Repeating the preceding analysis, where the functional Φ(w) replaces the term(w·w), an optimal solution (w,g,α) is found as before by maximizing a quadratic form, subject to the modified constraints
Σα_{j} ·y _{j}=0., (A-20a)
0≦α_{j}≦C. (A-20b)
Use of (only) hyperplanes in an input space is insufficient for certain classes of data. See the examples in
In a support vector machine, input vectors are mapped into a high dimension feature space Z through a selected nonlinear mapping. In the space Z, an optimal separating hyperplane is constructed that maximizes a certain Δ-margin associated with hyperplane separation.
First, consider a mapping that allows one to construct decision polynomials of degree 2 in the input space. One creates a (quadratic) feature space Z having dimension M=N(N+3)/2, with coordinates
u _{j} =x ^{j}(j=1, . . . , N: N coordinates) (A-21a)
u _{j+N} =x _{j} ^{2 }(j=1, . . . , N; N coordinates) (A-21b)
u _{j+2N} =x _{1}·_{2} , ·x _{1} ·x _{3} , . . . , x _{N-1} ·x _{N}, (N(N−1)/2 coordinates). (A-21c)
A separating hyperplane constructed in the space Z is assumed to be a second degree polynomial in the input space coordinates x_{j}(j=1, . . . N).
By analogy, in order to construct a polynomial of degree k in the input coordinates, one must construct a space Z having of the order of N^{k }coordinates, where one constructs an optimal separating hyperplane. For example, for k=4, the maximum number of coordinates needed in the space Z is
which is about 10^{8 }coordinates for a modest size input space of N=100 independent coordinates.
For a quadratic feature space Z, one first determines a kernel function K of inner products according to
(u _{L1} ·u _{L2})=K(x _{j1} ,x _{j2})=K(x _{j2} ,x _{j1}) (L1, L2=1, . . . ,N(N+3)/2). (A-23)
One constructs nonlinear decision functions
I(x)=sgn{Σα_{j} ·K(x,x _{j})+b0} (A-24)
Mercer (1909) has proved that a one-to-one correspondence exists between the set of symmetric, positive definite functions κ(x,y) defined on the real line that satisfy
∫∫κ(x,y) f(x) f(y) dx dy≧0 (A-27)
for any L2-integrable function f(x) satisfying
∫f(x)^{2 } dx <∞ (A-28)
and the set of inner products defined on that function space {f}. Thus, any kernel function K(x_{j1},x_{j2}) satisfying conditions of the Mercer theorem can be used to construct an inner product of the type set forth in Eq. (A-23). Using different expressions for the kernel K(x_{j1},x_{j2}), one can construct different learning machines with corresponding nonlinear decision functions.
For example, the kernel function
K(x′,x″)={(x′·x″)+1}q, (A-29)
can be used to specify polynomials of degree up to q (preferably an integer).
Much of the preceding development is taken from V.N. Vapnik, “An Overview of Statistical Learning Theory”, IEEE Trans. Neural Networks, vol. 10 (1999), pp. 988–999. The present invention provides a hybrid approach in which the input layer and hidden layer(s) of an NN component are used to create a data-adaptive feature space for an SVM component. As indicated in the preceding, the combined NN/SVM analysis of the invention is not limited to the particular NN analysis or to the particular SVM analysis set forth in this Appendix.
Cited Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|
US4837798 | Jun 2, 1986 | Jun 6, 1989 | American Telephone And Telegraph Company | Communication system having unified messaging |
US4910765 | Sep 6, 1988 | Mar 20, 1990 | Ricoh Company, Ltd. | Communication terminal apparatus having a relaying function |
US5513126 | Oct 4, 1993 | Apr 30, 1996 | Xerox Corporation | Network having selectively accessible recipient prioritized communication channel profiles |
US5630060 | May 1, 1995 | May 13, 1997 | Canon Kabushiki Kaisha | Method and apparatus for delivering multi-media messages over different transmission media |
US5740231 | Sep 16, 1994 | Apr 14, 1998 | Octel Communications Corporation | Network-based multimedia communications and directory system and method of operation |
US5781186 | Feb 2, 1996 | Jul 14, 1998 | Lucent Technologies Inc. | Arrangement for specifying presentation of multimedia message components |
US5848415 | Dec 18, 1996 | Dec 8, 1998 | Unisys Corporation | Selective multiple protocol transport and dynamic format conversion in a multi-user network |
US5864870 | Dec 18, 1996 | Jan 26, 1999 | Unisys Corp. | Method for storing/retrieving files of various formats in an object database using a virtual multimedia file system |
US6034970 | Jul 2, 1997 | Mar 7, 2000 | Adaptive Micro Systems, Inc. | Intelligent messaging system and method for providing and updating a message using a communication device, such as a large character display |
US6073165 | Jul 29, 1997 | Jun 6, 2000 | Jfax Communications, Inc. | Filtering computer network messages directed to a user's e-mail box based on user defined filters, and forwarding a filtered message to the user's receiver |
US6119137 | Jan 30, 1997 | Sep 12, 2000 | Tumbleweed Communications Corp. | Distributed dynamic document conversion server |
US6157945 | Jul 1, 1998 | Dec 5, 2000 | Ricoh Company, Ltd. | Digital communication device and method including a routing function |
US6167441 | Nov 21, 1997 | Dec 26, 2000 | International Business Machines Corporation | Customization of web pages based on requester type |
US6212550 | Jan 21, 1997 | Apr 3, 2001 | Motorola, Inc. | Method and system in a client-server for automatically converting messages from a first format to a second format compatible with a message retrieving device |
US20010031076 * | Feb 1, 2001 | Oct 18, 2001 | Renato Campanini | Method and apparatus for the automatic detection of microcalcifications in digital signals of mammary tissue |
US20030040904 * | Aug 27, 2001 | Feb 27, 2003 | Nec Research Institute, Inc. | Extracting classifying data in music from an audio bitstream |
US20030078850 * | Sep 5, 2001 | Apr 24, 2003 | Eric Hartman | Electronic marketplace system and method using a support vector machine |
Reference | ||
---|---|---|
1 | * | J. A. K. Suykens et al, Recurrent Least Squares Support Vector Machines, Jul. 2000, IEEE, 1057-7122/00, 1109-1114. |
2 | M.M. Rai and N.K. Madavan, "Aerodynamic Design Using Neural Networks", AIAA Jour., vol. 38 (2000). | |
3 | * | Pascal Vincent et al, A Neural Support Vector Network Architecture with Adaptive Kernels, 2000, IEEE,0-7695-0619-4, 187-19. |
4 | V.N. Vapnik, "An Overview of Statistical Learning Theory", IEEE Trans. on Neural Networks, vol. 10 (1999) pp. 988-999. | |
5 | * | Vladimir N. Vapnik, An Overview of Statistical Learning Theory, 1999, IEEE, 1045-9227/99. |
Citing Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|
US7457788 * | Jun 10, 2004 | Nov 25, 2008 | Oracle International Corporation | Reducing number of computations in a neural network modeling several data sets |
US7899652 | Aug 31, 2007 | Mar 1, 2011 | Toyota Motor Engineering & Manufacturing North America, Inc. | Linear programming support vector regression with wavelet kernel |
US7899767 * | Mar 1, 2011 | Kabushiki Kaisha Toshiba | Probabilistic model generation method, apparatus, and program | |
US8438142 * | May 7, 2013 | Google Inc. | Suggesting and refining user input based on original user input | |
US8463591 * | Jun 11, 2013 | Google Inc. | Efficient polynomial mapping of data for use with linear support vector machines | |
US8738271 | Dec 16, 2011 | May 27, 2014 | Toyota Motor Engineering & Manufacturing North America, Inc. | Asymmetric wavelet kernel in support vector learning |
US9020924 | Sep 13, 2012 | Apr 28, 2015 | Google Inc. | Suggesting and refining user input based on original user input |
US9235655 * | May 21, 2004 | Jan 12, 2016 | Hewlett-Packard Development Company, L.P. | Task-based design evaluation |
US20050261881 * | May 21, 2004 | Nov 24, 2005 | Hewlett-Packard Development Company, L.P. | Task-based design evaluation |
US20050278269 * | Jun 10, 2004 | Dec 15, 2005 | Oracle International Corporation | Reducing Number of Computations in a Neural Network Modeling Several Data Sets |
US20060253427 * | May 4, 2005 | Nov 9, 2006 | Jun Wu | Suggesting and refining user input based on original user input |
US20070038558 * | Aug 1, 2006 | Feb 15, 2007 | Kabushiki Kaisha Toshiba | Probabilistic model generation method, apparatus, and program |
US20080154816 * | Oct 31, 2006 | Jun 26, 2008 | Motorola, Inc. | Artificial neural network with adaptable infinite-logic nodes |
US20090063115 * | Aug 31, 2007 | Mar 5, 2009 | Zhao Lu | Linear programming support vector regression with wavelet kernel |
CN103488847A * | Oct 8, 2013 | Jan 1, 2014 | 北京航天长征飞行器研究所 | Aerodynamic shape optimization method based on neural network integration |
U.S. Classification | 706/21, 706/25, 706/19 |
International Classification | G06E1/00, G06E3/00, G06G7/00, G06F15/18, G06N3/04, G06K9/62 |
Cooperative Classification | G06K9/6269, G06N3/0454 |
European Classification | G06N3/04M, G06K9/62C1B |
Date | Code | Event | Description |
---|---|---|---|
Jul 1, 2005 | AS | Assignment | Owner name: USA AS REPRESENTED BY THE ADMINISTRATOR OF THE NAS Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:RAI, MAN MOHAN;REEL/FRAME:016463/0313 Effective date: 20020107 |
May 11, 2009 | REMI | Maintenance fee reminder mailed | |
Sep 24, 2009 | SULP | Surcharge for late payment | |
Sep 24, 2009 | FPAY | Fee payment | Year of fee payment: 4 |
Jun 14, 2013 | REMI | Maintenance fee reminder mailed | |
Nov 1, 2013 | LAPS | Lapse for failure to pay maintenance fees | |
Dec 24, 2013 | FP | Expired due to failure to pay maintenance fee | Effective date: 20131101 |